NewsAI

Mistral Large 4 “Le Chonk” Preview Arrives: 1 Trillion Parameters, Native Multimodal, DeepSWE 61.7% Tops Western Open Weights, Weights Out This Month

French AI company Mistral AI released the public preview of Mistral Large 4 (ML4, nicknamed “le Chonk”) on October 6: a 1-trillion-parameter, 49B-active native multimodal model positioned as the company’s largest and most capable to date. The preview API is live in Mistral Studio now, with model weights due out by the end of the month; until then, Mistral is red-teaming the model in real-world scenarios with outside experts and government agencies before release.

Key takeaways

  • Mistral Large 4 public preview: 1T total parameters, 49B active, natively multimodal; weights out this month
  • Trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own European data centers
  • DeepSWE v1.1 at 61.7% and Terminal-Bench 4 at 28.3%; composite Coding Agent Index 49.8%, ahead of DeepSeek V4 Pro and Qwen3.8 Max per the company
  • Standout cybersecurity showing: 93% on Cybench, 82% on the vulnerability-reproduction test; with reduced refusals for security research
  • Training data spans 160+ languages including every official EU language; human evals rank its coding quality second among five models

Positioning: a new frontier for Western open weights

Officially, ML4 is described as competitive with the strongest open-source models globally and far ahead of every open-weight model developed in the US and Europe; the company calls it state of the art among open models on enterprise-critical workloads like cybersecurity, finance, and law, and says it beats some closed frontier models on visual grounding. Training ran from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own European data centers, and the preview API runs on the same infrastructure; part of the “AI sovereignty” pitch, with European end-to-end deployment under European law.

Cybersecurity: a defensive tool with fewer refusals

ML4 ranks in the global top five on Artificial Analysis’ Cyber Index. On the test that asks a model to reproduce a real vulnerability in open-source software and then patch it, it scored 82%; the highest of any model, and solved 93% of Cybench’s 40 CTF-style challenges. The company notes that some closed models, including Claude Opus 5.5 and GPT-6 Astra, score near zero because safety filters refuse the task, even though defensive work often requires proving a vulnerability is real first. ML4’s design permits this kind of security research while still posting the highest refusal rate on malicious requests among open models (combined JailbreakBench, StrongREJECT, and AgentHarm), and it blocks 93.3% of indirect prompt-injection attacks on Lakera’s B3 benchmark. For security teams that need sovereign, auditable AI, the model will run in private cloud or on-premises deployments.

Coding and agentic workflows

On agentic coding, ML4 scores 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4, for a composite Coding Agent Index of 49.8%, which the company places ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. In blind human evals run with Surge AI, ML4 Preview ranked second of five models (3.74), behind Claude Opus 5 (4.22) and ahead of Kimi K3 (3.59) and GLM-5.3 (3.60). For agentic workflows, it scores 59.9% on AutomationBench’s 657 enterprise processes and 1,393 Elo on AA-Briefcase, a long-horizon knowledge-work benchmark.

Multimodality and science

Multimodality is the headline upgrade: the company calls ML4 one of the most capable models it has tested on visual grounding, scoring 42% on Dense 200 versus GPT-6 Astra’s 41%, with use cases from gigapixel satellite imagery analysis for disaster response to engineering-drawing inspection. On science, ML4 officially leads SciCode-Verified among open-weight models and demonstrated a one-shot, complete Hartree-Fock chemistry simulation; blind human evals favor it over GLM-5.3 on CAD and STEM tasks.

Enterprise workloads and independent checks

Law and finance are the verticals Mistral emphasizes most: on third-party Vals.ai evaluations, ML4 beats every open model on the HarveyAI Legal Agent benchmark and scores 67% on the Finch financial workflows benchmark, ahead of GPT-6 Astra on both. Two caveats deserve attention, though. First, the independent Artificial Analysis Intelligence Index currently gives the ML4 preview 38.4 points, a huge leap from Mistral Medium 3.5 (14) and Large 3 (9), and ahead of GLM-5.2 and DeepSeek V4 Pro, but still more than 19 points behind the frontier and eighth among open models. Second, some competitor comparisons in the official charts could not yet be independently verified from public sources; the picture will firm up once third parties can run the weights after this month’s release.

What comes next

Mistral says the weight release will include the model architecture, additional benchmarks, and details of its post-training approach, and that ML4 forms the base for a new generation of specialized and optimized Mistral models. Training data spans more than 160 languages, including every official EU language, and the model uses the same training, customization, and RL environments Mistral offers enterprise customers through Mistral Forge.

Sources: Mistral AI Blog, Trending Topics